Live demo

An agentic ops assistant you can run.

A small, anonymized version of the agentic systems I build at work. Give it an onboarding task; it looks the vendor up, runs compliance checks on its own, then stops and asks you to approve the one action that writes to a system of record. Tools and data are synthetic — the design is the point.

Pick a task and run the agent. It will look up the vendor, run compliance checks automatically, then pause for your approval before writing anything.

Why it's built this way

Agents that finish the job
The agent plans and executes end-to-end — look up, verify, then act — instead of just answering a question.
Human-in-the-loop where it counts
Read tools run automatically; the one write action pauses for explicit approval. Safety lives at the action boundary.
Observable by design
Every tool call, input, and result is surfaced — the same trace I'd instrument for evals and debugging in production.

Under the hood: a manual tool-use loop on the Claude API (claude-haiku-4-5), with the write tool gated for human approval. Without an API key it runs a faithful simulation so the demo always works; set ANTHROPIC_API_KEY for live model reasoning.

Kept cheap & safe: the model only ever runs from a fixed scenario (so it can't be used as a free LLM proxy), the approval step is fully server-side, and the endpoint is rate-limited and same-origin only.